| // NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --function-signature --include-generated-funcs --replace-function-regex "__omp_offloading_[0-9a-z]+_[0-9a-z]+" --prefix-filecheck-ir-name _ |
| // RUN: %clang_cc1 -DCK1 -verify -fopenmp -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=nvptx64-nvidia-cuda -fopenmp-cuda-mode -emit-llvm-bc %s -o %t-ppc-host.bc -fopenmp-version=45 |
| // RUN: %clang_cc1 -DCK1 -verify -fopenmp -x c++ -triple nvptx64-unknown-unknown -fopenmp-targets=nvptx64-nvidia-cuda -fopenmp-cuda-mode -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-ppc-host.bc -o - -debug-info-kind=limited -fopenmp-version=45 | FileCheck %s |
| // expected-no-diagnostics |
| |
| template <unsigned *ddd> |
| struct S { |
| static int a; |
| }; |
| |
| extern unsigned aaa; |
| template<> int S<&aaa>::a; |
| |
| template struct S<&aaa>; |
| |
| int main() { |
| /* int(*b)[a]; */ |
| /* int *(**c)[a]; */ |
| bool bb; |
| int a; |
| int b[10][10]; |
| int c[10][10][10]; |
| #pragma omp target parallel firstprivate(a, b) map(tofrom \ |
| : c) map(tofrom \ |
| : bb) if (target:a) |
| { |
| int &f = c[1][1][1]; |
| int &g = a; |
| int &h = b[1][1]; |
| int d = 15; |
| a = 5; |
| b[0][a] = 10; |
| c[0][0][a] = 11; |
| b[0][a] = c[0][0][a]; |
| bb |= b[0][a]; |
| } |
| #pragma omp target parallel firstprivate(a) map(tofrom \ |
| : c, b) map(to \ |
| : bb) |
| { |
| int &f = c[1][1][1]; |
| int &g = a; |
| int &h = b[1][1]; |
| int d = 15; |
| a = 5; |
| b[0][a] = 10; |
| c[0][0][a] = 11; |
| b[0][a] = c[0][0][a]; |
| d = bb; |
| } |
| #pragma omp target parallel map(tofrom \ |
| : a, c, b) map(from \ |
| : bb) |
| { |
| int &f = c[1][1][1]; |
| int &g = a; |
| int &h = b[1][1]; |
| int d = 15; |
| a = 5; |
| b[0][a] = 10; |
| c[0][0][a] = 11; |
| b[0][a] = c[0][0][a]; |
| bb = b[0][a]; |
| } |
| return 0; |
| } |
| // CHECK-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l23_debug__ |
| // CHECK-SAME: ([10 x [10 x [10 x i32]]] addrspace(1)* noalias [[C:%.*]], i32 [[A:%.*]], [10 x [10 x i32]]* noalias [[B:%.*]], i8 addrspace(1)* noalias [[BB:%.*]]) #[[ATTR0:[0-9]+]] !dbg [[DBG22:![0-9]+]] { |
| // CHECK-NEXT: entry: |
| // CHECK-NEXT: [[C_ADDR:%.*]] = alloca [10 x [10 x [10 x i32]]] addrspace(1)*, align 8 |
| // CHECK-NEXT: [[A_ADDR:%.*]] = alloca i32, align 4 |
| // CHECK-NEXT: [[B_ADDR:%.*]] = alloca [10 x [10 x i32]]*, align 8 |
| // CHECK-NEXT: [[BB_ADDR:%.*]] = alloca i8 addrspace(1)*, align 8 |
| // CHECK-NEXT: [[TMP:%.*]] = alloca [10 x [10 x [10 x i32]]]*, align 8 |
| // CHECK-NEXT: [[_TMP1:%.*]] = alloca [10 x [10 x i32]]*, align 8 |
| // CHECK-NEXT: [[_TMP2:%.*]] = alloca i8*, align 8 |
| // CHECK-NEXT: [[A_CASTED:%.*]] = alloca i64, align 8 |
| // CHECK-NEXT: [[CAPTURED_VARS_ADDRS:%.*]] = alloca [4 x i8*], align 8 |
| // CHECK-NEXT: store [10 x [10 x [10 x i32]]] addrspace(1)* [[C]], [10 x [10 x [10 x i32]]] addrspace(1)** [[C_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x [10 x i32]]] addrspace(1)** [[C_ADDR]], metadata [[META37:![0-9]+]], metadata !DIExpression()), !dbg [[DBG38:![0-9]+]] |
| // CHECK-NEXT: store i32 [[A]], i32* [[A_ADDR]], align 4 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32* [[A_ADDR]], metadata [[META39:![0-9]+]], metadata !DIExpression()), !dbg [[DBG40:![0-9]+]] |
| // CHECK-NEXT: store [10 x [10 x i32]]* [[B]], [10 x [10 x i32]]** [[B_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x i32]]** [[B_ADDR]], metadata [[META41:![0-9]+]], metadata !DIExpression()), !dbg [[DBG42:![0-9]+]] |
| // CHECK-NEXT: store i8 addrspace(1)* [[BB]], i8 addrspace(1)** [[BB_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i8 addrspace(1)** [[BB_ADDR]], metadata [[META43:![0-9]+]], metadata !DIExpression()), !dbg [[DBG44:![0-9]+]] |
| // CHECK-NEXT: [[TMP0:%.*]] = load [10 x [10 x [10 x i32]]] addrspace(1)*, [10 x [10 x [10 x i32]]] addrspace(1)** [[C_ADDR]], align 8, !dbg [[DBG45:![0-9]+]] |
| // CHECK-NEXT: [[TMP1:%.*]] = addrspacecast [10 x [10 x [10 x i32]]] addrspace(1)* [[TMP0]] to [10 x [10 x [10 x i32]]]*, !dbg [[DBG45]] |
| // CHECK-NEXT: store [10 x [10 x [10 x i32]]]* [[TMP1]], [10 x [10 x [10 x i32]]]** [[TMP]], align 8, !dbg [[DBG45]] |
| // CHECK-NEXT: [[TMP2:%.*]] = load [10 x [10 x [10 x i32]]]*, [10 x [10 x [10 x i32]]]** [[TMP]], align 8, !dbg [[DBG45]] |
| // CHECK-NEXT: [[TMP3:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[B_ADDR]], align 8, !dbg [[DBG45]] |
| // CHECK-NEXT: store [10 x [10 x i32]]* [[TMP3]], [10 x [10 x i32]]** [[_TMP1]], align 8, !dbg [[DBG45]] |
| // CHECK-NEXT: [[TMP4:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[_TMP1]], align 8, !dbg [[DBG45]] |
| // CHECK-NEXT: [[TMP5:%.*]] = load i8 addrspace(1)*, i8 addrspace(1)** [[BB_ADDR]], align 8, !dbg [[DBG45]] |
| // CHECK-NEXT: [[TMP6:%.*]] = addrspacecast i8 addrspace(1)* [[TMP5]] to i8*, !dbg [[DBG45]] |
| // CHECK-NEXT: store i8* [[TMP6]], i8** [[_TMP2]], align 8, !dbg [[DBG45]] |
| // CHECK-NEXT: [[TMP7:%.*]] = load i8*, i8** [[_TMP2]], align 8, !dbg [[DBG45]] |
| // CHECK-NEXT: [[NVPTX_NUM_THREADS:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.ntid.x(), !dbg [[DBG45]] |
| // CHECK-NEXT: call void @__kmpc_spmd_kernel_init(i32 [[NVPTX_NUM_THREADS]], i16 1), !dbg [[DBG45]] |
| // CHECK-NEXT: call void @__kmpc_data_sharing_init_stack_spmd(), !dbg [[DBG45]] |
| // CHECK-NEXT: br label [[DOTEXECUTE:%.*]], !dbg [[DBG45]] |
| // CHECK: .execute: |
| // CHECK-NEXT: [[TMP8:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB1:[0-9]+]]) |
| // CHECK-NEXT: [[TMP9:%.*]] = load i32, i32* [[A_ADDR]], align 4, !dbg [[DBG46:![0-9]+]] |
| // CHECK-NEXT: [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i32*, !dbg [[DBG46]] |
| // CHECK-NEXT: store i32 [[TMP9]], i32* [[CONV]], align 4, !dbg [[DBG46]] |
| // CHECK-NEXT: [[TMP10:%.*]] = load i64, i64* [[A_CASTED]], align 8, !dbg [[DBG46]] |
| // CHECK-NEXT: [[TMP11:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 0, !dbg [[DBG46]] |
| // CHECK-NEXT: [[TMP12:%.*]] = bitcast [10 x [10 x [10 x i32]]]* [[TMP2]] to i8*, !dbg [[DBG46]] |
| // CHECK-NEXT: store i8* [[TMP12]], i8** [[TMP11]], align 8, !dbg [[DBG46]] |
| // CHECK-NEXT: [[TMP13:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 1, !dbg [[DBG46]] |
| // CHECK-NEXT: [[TMP14:%.*]] = inttoptr i64 [[TMP10]] to i8*, !dbg [[DBG46]] |
| // CHECK-NEXT: store i8* [[TMP14]], i8** [[TMP13]], align 8, !dbg [[DBG46]] |
| // CHECK-NEXT: [[TMP15:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 2, !dbg [[DBG46]] |
| // CHECK-NEXT: [[TMP16:%.*]] = bitcast [10 x [10 x i32]]* [[TMP4]] to i8*, !dbg [[DBG46]] |
| // CHECK-NEXT: store i8* [[TMP16]], i8** [[TMP15]], align 8, !dbg [[DBG46]] |
| // CHECK-NEXT: [[TMP17:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 3, !dbg [[DBG46]] |
| // CHECK-NEXT: store i8* [[TMP7]], i8** [[TMP17]], align 8, !dbg [[DBG46]] |
| // CHECK-NEXT: [[TMP18:%.*]] = bitcast [4 x i8*]* [[CAPTURED_VARS_ADDRS]] to i8**, !dbg [[DBG46]] |
| // CHECK-NEXT: call void @__kmpc_parallel_51(%struct.ident_t* @[[GLOB1]], i32 [[TMP8]], i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*, [10 x [10 x [10 x i32]]]*, i64, [10 x [10 x i32]]*, i8*)* @__omp_outlined__ to i8*), i8* null, i8** [[TMP18]], i64 4), !dbg [[DBG46]] |
| // CHECK-NEXT: br label [[DOTOMP_DEINIT:%.*]], !dbg [[DBG47:![0-9]+]] |
| // CHECK: .omp.deinit: |
| // CHECK-NEXT: call void @__kmpc_spmd_kernel_deinit_v2(i16 1), !dbg [[DBG47]] |
| // CHECK-NEXT: br label [[DOTEXIT:%.*]], !dbg [[DBG47]] |
| // CHECK: .exit: |
| // CHECK-NEXT: ret void, !dbg [[DBG49:![0-9]+]] |
| // |
| // |
| // CHECK-LABEL: define {{[^@]+}}@__omp_outlined___debug__ |
| // CHECK-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], [10 x [10 x [10 x i32]]] addrspace(1)* noalias [[C:%.*]], i32 [[A:%.*]], [10 x [10 x i32]]* noalias [[B:%.*]], i8 addrspace(1)* noalias [[BB:%.*]]) #[[ATTR0]] !dbg [[DBG50:![0-9]+]] { |
| // CHECK-NEXT: entry: |
| // CHECK-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[C_ADDR:%.*]] = alloca [10 x [10 x [10 x i32]]] addrspace(1)*, align 8 |
| // CHECK-NEXT: [[A_ADDR:%.*]] = alloca i32, align 4 |
| // CHECK-NEXT: [[B_ADDR:%.*]] = alloca [10 x [10 x i32]]*, align 8 |
| // CHECK-NEXT: [[BB_ADDR:%.*]] = alloca i8 addrspace(1)*, align 8 |
| // CHECK-NEXT: [[TMP:%.*]] = alloca [10 x [10 x [10 x i32]]]*, align 8 |
| // CHECK-NEXT: [[_TMP1:%.*]] = alloca [10 x [10 x i32]]*, align 8 |
| // CHECK-NEXT: [[_TMP2:%.*]] = alloca i8*, align 8 |
| // CHECK-NEXT: [[B3:%.*]] = alloca [10 x [10 x i32]], align 4 |
| // CHECK-NEXT: [[F:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[G:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[H:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[D:%.*]] = alloca i32, align 4 |
| // CHECK-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[DOTGLOBAL_TID__ADDR]], metadata [[META57:![0-9]+]], metadata !DIExpression()), !dbg [[DBG58:![0-9]+]] |
| // CHECK-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[DOTBOUND_TID__ADDR]], metadata [[META59:![0-9]+]], metadata !DIExpression()), !dbg [[DBG58]] |
| // CHECK-NEXT: store [10 x [10 x [10 x i32]]] addrspace(1)* [[C]], [10 x [10 x [10 x i32]]] addrspace(1)** [[C_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x [10 x i32]]] addrspace(1)** [[C_ADDR]], metadata [[META60:![0-9]+]], metadata !DIExpression()), !dbg [[DBG61:![0-9]+]] |
| // CHECK-NEXT: store i32 [[A]], i32* [[A_ADDR]], align 4 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32* [[A_ADDR]], metadata [[META62:![0-9]+]], metadata !DIExpression()), !dbg [[DBG63:![0-9]+]] |
| // CHECK-NEXT: store [10 x [10 x i32]]* [[B]], [10 x [10 x i32]]** [[B_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x i32]]** [[B_ADDR]], metadata [[META64:![0-9]+]], metadata !DIExpression()), !dbg [[DBG65:![0-9]+]] |
| // CHECK-NEXT: store i8 addrspace(1)* [[BB]], i8 addrspace(1)** [[BB_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i8 addrspace(1)** [[BB_ADDR]], metadata [[META66:![0-9]+]], metadata !DIExpression()), !dbg [[DBG67:![0-9]+]] |
| // CHECK-NEXT: [[TMP0:%.*]] = load [10 x [10 x [10 x i32]]] addrspace(1)*, [10 x [10 x [10 x i32]]] addrspace(1)** [[C_ADDR]], align 8, !dbg [[DBG68:![0-9]+]] |
| // CHECK-NEXT: [[TMP1:%.*]] = addrspacecast [10 x [10 x [10 x i32]]] addrspace(1)* [[TMP0]] to [10 x [10 x [10 x i32]]]*, !dbg [[DBG68]] |
| // CHECK-NEXT: store [10 x [10 x [10 x i32]]]* [[TMP1]], [10 x [10 x [10 x i32]]]** [[TMP]], align 8, !dbg [[DBG68]] |
| // CHECK-NEXT: [[TMP2:%.*]] = load [10 x [10 x [10 x i32]]]*, [10 x [10 x [10 x i32]]]** [[TMP]], align 8, !dbg [[DBG68]] |
| // CHECK-NEXT: [[TMP3:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[B_ADDR]], align 8, !dbg [[DBG68]] |
| // CHECK-NEXT: store [10 x [10 x i32]]* [[TMP3]], [10 x [10 x i32]]** [[_TMP1]], align 8, !dbg [[DBG68]] |
| // CHECK-NEXT: [[TMP4:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[_TMP1]], align 8, !dbg [[DBG68]] |
| // CHECK-NEXT: [[TMP5:%.*]] = load i8 addrspace(1)*, i8 addrspace(1)** [[BB_ADDR]], align 8, !dbg [[DBG68]] |
| // CHECK-NEXT: [[TMP6:%.*]] = addrspacecast i8 addrspace(1)* [[TMP5]] to i8*, !dbg [[DBG68]] |
| // CHECK-NEXT: store i8* [[TMP6]], i8** [[_TMP2]], align 8, !dbg [[DBG68]] |
| // CHECK-NEXT: [[TMP7:%.*]] = load i8*, i8** [[_TMP2]], align 8, !dbg [[DBG68]] |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x i32]]* [[B3]], metadata [[META69:![0-9]+]], metadata !DIExpression()), !dbg [[DBG58]] |
| // CHECK-NEXT: [[TMP8:%.*]] = bitcast [10 x [10 x i32]]* [[B3]] to i8*, !dbg [[DBG68]] |
| // CHECK-NEXT: [[TMP9:%.*]] = bitcast [10 x [10 x i32]]* [[TMP4]] to i8*, !dbg [[DBG68]] |
| // CHECK-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[TMP8]], i8* align 4 [[TMP9]], i64 400, i1 false), !dbg [[DBG68]] |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[F]], metadata [[META70:![0-9]+]], metadata !DIExpression()), !dbg [[DBG73:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x [10 x [10 x i32]]], [10 x [10 x [10 x i32]]]* [[TMP2]], i64 0, i64 1, !dbg [[DBG74:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX4:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[ARRAYIDX]], i64 0, i64 1, !dbg [[DBG74]] |
| // CHECK-NEXT: [[ARRAYIDX5:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX4]], i64 0, i64 1, !dbg [[DBG74]] |
| // CHECK-NEXT: store i32* [[ARRAYIDX5]], i32** [[F]], align 8, !dbg [[DBG73]] |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[G]], metadata [[META75:![0-9]+]], metadata !DIExpression()), !dbg [[DBG76:![0-9]+]] |
| // CHECK-NEXT: store i32* [[A_ADDR]], i32** [[G]], align 8, !dbg [[DBG76]] |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[H]], metadata [[META77:![0-9]+]], metadata !DIExpression()), !dbg [[DBG78:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX6:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[B3]], i64 0, i64 1, !dbg [[DBG79:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX7:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX6]], i64 0, i64 1, !dbg [[DBG79]] |
| // CHECK-NEXT: store i32* [[ARRAYIDX7]], i32** [[H]], align 8, !dbg [[DBG78]] |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32* [[D]], metadata [[META80:![0-9]+]], metadata !DIExpression()), !dbg [[DBG81:![0-9]+]] |
| // CHECK-NEXT: store i32 15, i32* [[D]], align 4, !dbg [[DBG81]] |
| // CHECK-NEXT: store i32 5, i32* [[A_ADDR]], align 4, !dbg [[DBG82:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX8:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[B3]], i64 0, i64 0, !dbg [[DBG83:![0-9]+]] |
| // CHECK-NEXT: [[TMP10:%.*]] = load i32, i32* [[A_ADDR]], align 4, !dbg [[DBG84:![0-9]+]] |
| // CHECK-NEXT: [[IDXPROM:%.*]] = sext i32 [[TMP10]] to i64, !dbg [[DBG83]] |
| // CHECK-NEXT: [[ARRAYIDX9:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX8]], i64 0, i64 [[IDXPROM]], !dbg [[DBG83]] |
| // CHECK-NEXT: store i32 10, i32* [[ARRAYIDX9]], align 4, !dbg [[DBG85:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX10:%.*]] = getelementptr inbounds [10 x [10 x [10 x i32]]], [10 x [10 x [10 x i32]]]* [[TMP2]], i64 0, i64 0, !dbg [[DBG86:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX11:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[ARRAYIDX10]], i64 0, i64 0, !dbg [[DBG86]] |
| // CHECK-NEXT: [[TMP11:%.*]] = load i32, i32* [[A_ADDR]], align 4, !dbg [[DBG87:![0-9]+]] |
| // CHECK-NEXT: [[IDXPROM12:%.*]] = sext i32 [[TMP11]] to i64, !dbg [[DBG86]] |
| // CHECK-NEXT: [[ARRAYIDX13:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX11]], i64 0, i64 [[IDXPROM12]], !dbg [[DBG86]] |
| // CHECK-NEXT: store i32 11, i32* [[ARRAYIDX13]], align 4, !dbg [[DBG88:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX14:%.*]] = getelementptr inbounds [10 x [10 x [10 x i32]]], [10 x [10 x [10 x i32]]]* [[TMP2]], i64 0, i64 0, !dbg [[DBG89:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX15:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[ARRAYIDX14]], i64 0, i64 0, !dbg [[DBG89]] |
| // CHECK-NEXT: [[TMP12:%.*]] = load i32, i32* [[A_ADDR]], align 4, !dbg [[DBG90:![0-9]+]] |
| // CHECK-NEXT: [[IDXPROM16:%.*]] = sext i32 [[TMP12]] to i64, !dbg [[DBG89]] |
| // CHECK-NEXT: [[ARRAYIDX17:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX15]], i64 0, i64 [[IDXPROM16]], !dbg [[DBG89]] |
| // CHECK-NEXT: [[TMP13:%.*]] = load i32, i32* [[ARRAYIDX17]], align 4, !dbg [[DBG89]] |
| // CHECK-NEXT: [[ARRAYIDX18:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[B3]], i64 0, i64 0, !dbg [[DBG91:![0-9]+]] |
| // CHECK-NEXT: [[TMP14:%.*]] = load i32, i32* [[A_ADDR]], align 4, !dbg [[DBG92:![0-9]+]] |
| // CHECK-NEXT: [[IDXPROM19:%.*]] = sext i32 [[TMP14]] to i64, !dbg [[DBG91]] |
| // CHECK-NEXT: [[ARRAYIDX20:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX18]], i64 0, i64 [[IDXPROM19]], !dbg [[DBG91]] |
| // CHECK-NEXT: store i32 [[TMP13]], i32* [[ARRAYIDX20]], align 4, !dbg [[DBG93:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX21:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[B3]], i64 0, i64 0, !dbg [[DBG94:![0-9]+]] |
| // CHECK-NEXT: [[TMP15:%.*]] = load i32, i32* [[A_ADDR]], align 4, !dbg [[DBG95:![0-9]+]] |
| // CHECK-NEXT: [[IDXPROM22:%.*]] = sext i32 [[TMP15]] to i64, !dbg [[DBG94]] |
| // CHECK-NEXT: [[ARRAYIDX23:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX21]], i64 0, i64 [[IDXPROM22]], !dbg [[DBG94]] |
| // CHECK-NEXT: [[TMP16:%.*]] = load i32, i32* [[ARRAYIDX23]], align 4, !dbg [[DBG94]] |
| // CHECK-NEXT: [[TMP17:%.*]] = load i8, i8* [[TMP7]], align 1, !dbg [[DBG96:![0-9]+]] |
| // CHECK-NEXT: [[TOBOOL:%.*]] = trunc i8 [[TMP17]] to i1, !dbg [[DBG96]] |
| // CHECK-NEXT: [[CONV:%.*]] = zext i1 [[TOBOOL]] to i32, !dbg [[DBG96]] |
| // CHECK-NEXT: [[OR:%.*]] = or i32 [[CONV]], [[TMP16]], !dbg [[DBG96]] |
| // CHECK-NEXT: [[TOBOOL24:%.*]] = icmp ne i32 [[OR]], 0, !dbg [[DBG96]] |
| // CHECK-NEXT: [[FROMBOOL:%.*]] = zext i1 [[TOBOOL24]] to i8, !dbg [[DBG96]] |
| // CHECK-NEXT: store i8 [[FROMBOOL]], i8* [[TMP7]], align 1, !dbg [[DBG96]] |
| // CHECK-NEXT: ret void, !dbg [[DBG97:![0-9]+]] |
| // |
| // |
| // CHECK-LABEL: define {{[^@]+}}@__omp_outlined__ |
| // CHECK-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], [10 x [10 x [10 x i32]]]* nonnull align 4 dereferenceable(4000) [[C:%.*]], i64 [[A:%.*]], [10 x [10 x i32]]* nonnull align 4 dereferenceable(400) [[B:%.*]], i8* nonnull align 1 dereferenceable(1) [[BB:%.*]]) #[[ATTR0]] !dbg [[DBG98:![0-9]+]] { |
| // CHECK-NEXT: entry: |
| // CHECK-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[C_ADDR:%.*]] = alloca [10 x [10 x [10 x i32]]]*, align 8 |
| // CHECK-NEXT: [[A_ADDR:%.*]] = alloca i64, align 8 |
| // CHECK-NEXT: [[B_ADDR:%.*]] = alloca [10 x [10 x i32]]*, align 8 |
| // CHECK-NEXT: [[BB_ADDR:%.*]] = alloca i8*, align 8 |
| // CHECK-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[DOTGLOBAL_TID__ADDR]], metadata [[META105:![0-9]+]], metadata !DIExpression()), !dbg [[DBG106:![0-9]+]] |
| // CHECK-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[DOTBOUND_TID__ADDR]], metadata [[META107:![0-9]+]], metadata !DIExpression()), !dbg [[DBG106]] |
| // CHECK-NEXT: store [10 x [10 x [10 x i32]]]* [[C]], [10 x [10 x [10 x i32]]]** [[C_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x [10 x i32]]]** [[C_ADDR]], metadata [[META108:![0-9]+]], metadata !DIExpression()), !dbg [[DBG106]] |
| // CHECK-NEXT: store i64 [[A]], i64* [[A_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i64* [[A_ADDR]], metadata [[META109:![0-9]+]], metadata !DIExpression()), !dbg [[DBG106]] |
| // CHECK-NEXT: store [10 x [10 x i32]]* [[B]], [10 x [10 x i32]]** [[B_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x i32]]** [[B_ADDR]], metadata [[META110:![0-9]+]], metadata !DIExpression()), !dbg [[DBG106]] |
| // CHECK-NEXT: store i8* [[BB]], i8** [[BB_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i8** [[BB_ADDR]], metadata [[META111:![0-9]+]], metadata !DIExpression()), !dbg [[DBG106]] |
| // CHECK-NEXT: [[TMP0:%.*]] = load [10 x [10 x [10 x i32]]]*, [10 x [10 x [10 x i32]]]** [[C_ADDR]], align 8, !dbg [[DBG112:![0-9]+]] |
| // CHECK-NEXT: [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*, !dbg [[DBG112]] |
| // CHECK-NEXT: [[TMP1:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[B_ADDR]], align 8, !dbg [[DBG112]] |
| // CHECK-NEXT: [[TMP2:%.*]] = load i8*, i8** [[BB_ADDR]], align 8, !dbg [[DBG112]] |
| // CHECK-NEXT: [[TMP3:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8, !dbg [[DBG112]] |
| // CHECK-NEXT: [[TMP4:%.*]] = load i32*, i32** [[DOTBOUND_TID__ADDR]], align 8, !dbg [[DBG112]] |
| // CHECK-NEXT: [[TMP5:%.*]] = load [10 x [10 x [10 x i32]]]*, [10 x [10 x [10 x i32]]]** [[C_ADDR]], align 8, !dbg [[DBG112]] |
| // CHECK-NEXT: [[TMP6:%.*]] = load i32, i32* [[CONV]], align 8, !dbg [[DBG112]] |
| // CHECK-NEXT: [[TMP7:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[B_ADDR]], align 8, !dbg [[DBG112]] |
| // CHECK-NEXT: [[TMP8:%.*]] = load i8*, i8** [[BB_ADDR]], align 8, !dbg [[DBG112]] |
| // CHECK-NEXT: [[TMP9:%.*]] = addrspacecast [10 x [10 x [10 x i32]]]* [[TMP5]] to [10 x [10 x [10 x i32]]] addrspace(1)*, !dbg [[DBG112]] |
| // CHECK-NEXT: [[TMP10:%.*]] = addrspacecast i8* [[TMP8]] to i8 addrspace(1)*, !dbg [[DBG112]] |
| // CHECK-NEXT: call void @__omp_outlined___debug__(i32* [[TMP3]], i32* [[TMP4]], [10 x [10 x [10 x i32]]] addrspace(1)* [[TMP9]], i32 [[TMP6]], [10 x [10 x i32]]* [[TMP7]], i8 addrspace(1)* [[TMP10]]) #[[ATTR4:[0-9]+]], !dbg [[DBG112]] |
| // CHECK-NEXT: ret void, !dbg [[DBG112]] |
| // |
| // |
| // CHECK-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l23 |
| // CHECK-SAME: ([10 x [10 x [10 x i32]]]* nonnull align 4 dereferenceable(4000) [[C:%.*]], i64 [[A:%.*]], [10 x [10 x i32]]* nonnull align 4 dereferenceable(400) [[B:%.*]], i8* nonnull align 1 dereferenceable(1) [[BB:%.*]]) #[[ATTR0]] !dbg [[DBG113:![0-9]+]] { |
| // CHECK-NEXT: entry: |
| // CHECK-NEXT: [[C_ADDR:%.*]] = alloca [10 x [10 x [10 x i32]]]*, align 8 |
| // CHECK-NEXT: [[A_ADDR:%.*]] = alloca i64, align 8 |
| // CHECK-NEXT: [[B_ADDR:%.*]] = alloca [10 x [10 x i32]]*, align 8 |
| // CHECK-NEXT: [[BB_ADDR:%.*]] = alloca i8*, align 8 |
| // CHECK-NEXT: store [10 x [10 x [10 x i32]]]* [[C]], [10 x [10 x [10 x i32]]]** [[C_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x [10 x i32]]]** [[C_ADDR]], metadata [[META116:![0-9]+]], metadata !DIExpression()), !dbg [[DBG117:![0-9]+]] |
| // CHECK-NEXT: store i64 [[A]], i64* [[A_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i64* [[A_ADDR]], metadata [[META118:![0-9]+]], metadata !DIExpression()), !dbg [[DBG117]] |
| // CHECK-NEXT: store [10 x [10 x i32]]* [[B]], [10 x [10 x i32]]** [[B_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x i32]]** [[B_ADDR]], metadata [[META119:![0-9]+]], metadata !DIExpression()), !dbg [[DBG117]] |
| // CHECK-NEXT: store i8* [[BB]], i8** [[BB_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i8** [[BB_ADDR]], metadata [[META120:![0-9]+]], metadata !DIExpression()), !dbg [[DBG117]] |
| // CHECK-NEXT: [[TMP0:%.*]] = load [10 x [10 x [10 x i32]]]*, [10 x [10 x [10 x i32]]]** [[C_ADDR]], align 8, !dbg [[DBG121:![0-9]+]] |
| // CHECK-NEXT: [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*, !dbg [[DBG121]] |
| // CHECK-NEXT: [[TMP1:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[B_ADDR]], align 8, !dbg [[DBG121]] |
| // CHECK-NEXT: [[TMP2:%.*]] = load i8*, i8** [[BB_ADDR]], align 8, !dbg [[DBG121]] |
| // CHECK-NEXT: [[TMP3:%.*]] = load [10 x [10 x [10 x i32]]]*, [10 x [10 x [10 x i32]]]** [[C_ADDR]], align 8, !dbg [[DBG121]] |
| // CHECK-NEXT: [[TMP4:%.*]] = load i32, i32* [[CONV]], align 8, !dbg [[DBG121]] |
| // CHECK-NEXT: [[TMP5:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[B_ADDR]], align 8, !dbg [[DBG121]] |
| // CHECK-NEXT: [[TMP6:%.*]] = load i8*, i8** [[BB_ADDR]], align 8, !dbg [[DBG121]] |
| // CHECK-NEXT: [[TMP7:%.*]] = addrspacecast [10 x [10 x [10 x i32]]]* [[TMP3]] to [10 x [10 x [10 x i32]]] addrspace(1)*, !dbg [[DBG121]] |
| // CHECK-NEXT: [[TMP8:%.*]] = addrspacecast i8* [[TMP6]] to i8 addrspace(1)*, !dbg [[DBG121]] |
| // CHECK-NEXT: call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l23_debug__([10 x [10 x [10 x i32]]] addrspace(1)* [[TMP7]], i32 [[TMP4]], [10 x [10 x i32]]* [[TMP5]], i8 addrspace(1)* [[TMP8]]) #[[ATTR4]], !dbg [[DBG121]] |
| // CHECK-NEXT: ret void, !dbg [[DBG121]] |
| // |
| // |
| // CHECK-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l37_debug__ |
| // CHECK-SAME: ([10 x [10 x [10 x i32]]] addrspace(1)* noalias [[C:%.*]], i32 [[A:%.*]], [10 x [10 x i32]] addrspace(1)* noalias [[B:%.*]], i8 addrspace(1)* noalias [[BB:%.*]]) #[[ATTR0]] !dbg [[DBG122:![0-9]+]] { |
| // CHECK-NEXT: entry: |
| // CHECK-NEXT: [[C_ADDR:%.*]] = alloca [10 x [10 x [10 x i32]]] addrspace(1)*, align 8 |
| // CHECK-NEXT: [[A_ADDR:%.*]] = alloca i32, align 4 |
| // CHECK-NEXT: [[B_ADDR:%.*]] = alloca [10 x [10 x i32]] addrspace(1)*, align 8 |
| // CHECK-NEXT: [[BB_ADDR:%.*]] = alloca i8 addrspace(1)*, align 8 |
| // CHECK-NEXT: [[TMP:%.*]] = alloca [10 x [10 x [10 x i32]]]*, align 8 |
| // CHECK-NEXT: [[_TMP1:%.*]] = alloca [10 x [10 x i32]]*, align 8 |
| // CHECK-NEXT: [[_TMP2:%.*]] = alloca i8*, align 8 |
| // CHECK-NEXT: [[A_CASTED:%.*]] = alloca i64, align 8 |
| // CHECK-NEXT: [[CAPTURED_VARS_ADDRS:%.*]] = alloca [4 x i8*], align 8 |
| // CHECK-NEXT: store [10 x [10 x [10 x i32]]] addrspace(1)* [[C]], [10 x [10 x [10 x i32]]] addrspace(1)** [[C_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x [10 x i32]]] addrspace(1)** [[C_ADDR]], metadata [[META127:![0-9]+]], metadata !DIExpression()), !dbg [[DBG128:![0-9]+]] |
| // CHECK-NEXT: store i32 [[A]], i32* [[A_ADDR]], align 4 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32* [[A_ADDR]], metadata [[META129:![0-9]+]], metadata !DIExpression()), !dbg [[DBG130:![0-9]+]] |
| // CHECK-NEXT: store [10 x [10 x i32]] addrspace(1)* [[B]], [10 x [10 x i32]] addrspace(1)** [[B_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x i32]] addrspace(1)** [[B_ADDR]], metadata [[META131:![0-9]+]], metadata !DIExpression()), !dbg [[DBG132:![0-9]+]] |
| // CHECK-NEXT: store i8 addrspace(1)* [[BB]], i8 addrspace(1)** [[BB_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i8 addrspace(1)** [[BB_ADDR]], metadata [[META133:![0-9]+]], metadata !DIExpression()), !dbg [[DBG134:![0-9]+]] |
| // CHECK-NEXT: [[TMP0:%.*]] = load [10 x [10 x [10 x i32]]] addrspace(1)*, [10 x [10 x [10 x i32]]] addrspace(1)** [[C_ADDR]], align 8, !dbg [[DBG135:![0-9]+]] |
| // CHECK-NEXT: [[TMP1:%.*]] = addrspacecast [10 x [10 x [10 x i32]]] addrspace(1)* [[TMP0]] to [10 x [10 x [10 x i32]]]*, !dbg [[DBG135]] |
| // CHECK-NEXT: store [10 x [10 x [10 x i32]]]* [[TMP1]], [10 x [10 x [10 x i32]]]** [[TMP]], align 8, !dbg [[DBG135]] |
| // CHECK-NEXT: [[TMP2:%.*]] = load [10 x [10 x [10 x i32]]]*, [10 x [10 x [10 x i32]]]** [[TMP]], align 8, !dbg [[DBG135]] |
| // CHECK-NEXT: [[TMP3:%.*]] = load [10 x [10 x i32]] addrspace(1)*, [10 x [10 x i32]] addrspace(1)** [[B_ADDR]], align 8, !dbg [[DBG135]] |
| // CHECK-NEXT: [[TMP4:%.*]] = addrspacecast [10 x [10 x i32]] addrspace(1)* [[TMP3]] to [10 x [10 x i32]]*, !dbg [[DBG135]] |
| // CHECK-NEXT: store [10 x [10 x i32]]* [[TMP4]], [10 x [10 x i32]]** [[_TMP1]], align 8, !dbg [[DBG135]] |
| // CHECK-NEXT: [[TMP5:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[_TMP1]], align 8, !dbg [[DBG135]] |
| // CHECK-NEXT: [[TMP6:%.*]] = load i8 addrspace(1)*, i8 addrspace(1)** [[BB_ADDR]], align 8, !dbg [[DBG135]] |
| // CHECK-NEXT: [[TMP7:%.*]] = addrspacecast i8 addrspace(1)* [[TMP6]] to i8*, !dbg [[DBG135]] |
| // CHECK-NEXT: store i8* [[TMP7]], i8** [[_TMP2]], align 8, !dbg [[DBG135]] |
| // CHECK-NEXT: [[TMP8:%.*]] = load i8*, i8** [[_TMP2]], align 8, !dbg [[DBG135]] |
| // CHECK-NEXT: [[NVPTX_NUM_THREADS:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.ntid.x(), !dbg [[DBG135]] |
| // CHECK-NEXT: call void @__kmpc_spmd_kernel_init(i32 [[NVPTX_NUM_THREADS]], i16 1), !dbg [[DBG135]] |
| // CHECK-NEXT: call void @__kmpc_data_sharing_init_stack_spmd(), !dbg [[DBG135]] |
| // CHECK-NEXT: br label [[DOTEXECUTE:%.*]], !dbg [[DBG135]] |
| // CHECK: .execute: |
| // CHECK-NEXT: [[TMP9:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB3:[0-9]+]]) |
| // CHECK-NEXT: [[TMP10:%.*]] = load i32, i32* [[A_ADDR]], align 4, !dbg [[DBG136:![0-9]+]] |
| // CHECK-NEXT: [[CONV:%.*]] = bitcast i64* [[A_CASTED]] to i32*, !dbg [[DBG136]] |
| // CHECK-NEXT: store i32 [[TMP10]], i32* [[CONV]], align 4, !dbg [[DBG136]] |
| // CHECK-NEXT: [[TMP11:%.*]] = load i64, i64* [[A_CASTED]], align 8, !dbg [[DBG136]] |
| // CHECK-NEXT: [[TMP12:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 0, !dbg [[DBG136]] |
| // CHECK-NEXT: [[TMP13:%.*]] = bitcast [10 x [10 x [10 x i32]]]* [[TMP2]] to i8*, !dbg [[DBG136]] |
| // CHECK-NEXT: store i8* [[TMP13]], i8** [[TMP12]], align 8, !dbg [[DBG136]] |
| // CHECK-NEXT: [[TMP14:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 1, !dbg [[DBG136]] |
| // CHECK-NEXT: [[TMP15:%.*]] = inttoptr i64 [[TMP11]] to i8*, !dbg [[DBG136]] |
| // CHECK-NEXT: store i8* [[TMP15]], i8** [[TMP14]], align 8, !dbg [[DBG136]] |
| // CHECK-NEXT: [[TMP16:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 2, !dbg [[DBG136]] |
| // CHECK-NEXT: [[TMP17:%.*]] = bitcast [10 x [10 x i32]]* [[TMP5]] to i8*, !dbg [[DBG136]] |
| // CHECK-NEXT: store i8* [[TMP17]], i8** [[TMP16]], align 8, !dbg [[DBG136]] |
| // CHECK-NEXT: [[TMP18:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 3, !dbg [[DBG136]] |
| // CHECK-NEXT: store i8* [[TMP8]], i8** [[TMP18]], align 8, !dbg [[DBG136]] |
| // CHECK-NEXT: [[TMP19:%.*]] = bitcast [4 x i8*]* [[CAPTURED_VARS_ADDRS]] to i8**, !dbg [[DBG136]] |
| // CHECK-NEXT: call void @__kmpc_parallel_51(%struct.ident_t* @[[GLOB3]], i32 [[TMP9]], i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*, [10 x [10 x [10 x i32]]]*, i64, [10 x [10 x i32]]*, i8*)* @__omp_outlined__2 to i8*), i8* null, i8** [[TMP19]], i64 4), !dbg [[DBG136]] |
| // CHECK-NEXT: br label [[DOTOMP_DEINIT:%.*]], !dbg [[DBG137:![0-9]+]] |
| // CHECK: .omp.deinit: |
| // CHECK-NEXT: call void @__kmpc_spmd_kernel_deinit_v2(i16 1), !dbg [[DBG137]] |
| // CHECK-NEXT: br label [[DOTEXIT:%.*]], !dbg [[DBG137]] |
| // CHECK: .exit: |
| // CHECK-NEXT: ret void, !dbg [[DBG139:![0-9]+]] |
| // |
| // |
| // CHECK-LABEL: define {{[^@]+}}@__omp_outlined___debug__1 |
| // CHECK-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], [10 x [10 x [10 x i32]]] addrspace(1)* noalias [[C:%.*]], i32 [[A:%.*]], [10 x [10 x i32]] addrspace(1)* noalias [[B:%.*]], i8 addrspace(1)* noalias [[BB:%.*]]) #[[ATTR0]] !dbg [[DBG140:![0-9]+]] { |
| // CHECK-NEXT: entry: |
| // CHECK-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[C_ADDR:%.*]] = alloca [10 x [10 x [10 x i32]]] addrspace(1)*, align 8 |
| // CHECK-NEXT: [[A_ADDR:%.*]] = alloca i32, align 4 |
| // CHECK-NEXT: [[B_ADDR:%.*]] = alloca [10 x [10 x i32]] addrspace(1)*, align 8 |
| // CHECK-NEXT: [[BB_ADDR:%.*]] = alloca i8 addrspace(1)*, align 8 |
| // CHECK-NEXT: [[TMP:%.*]] = alloca [10 x [10 x [10 x i32]]]*, align 8 |
| // CHECK-NEXT: [[_TMP1:%.*]] = alloca [10 x [10 x i32]]*, align 8 |
| // CHECK-NEXT: [[_TMP2:%.*]] = alloca i8*, align 8 |
| // CHECK-NEXT: [[F:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[G:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[H:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[D:%.*]] = alloca i32, align 4 |
| // CHECK-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[DOTGLOBAL_TID__ADDR]], metadata [[META143:![0-9]+]], metadata !DIExpression()), !dbg [[DBG144:![0-9]+]] |
| // CHECK-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[DOTBOUND_TID__ADDR]], metadata [[META145:![0-9]+]], metadata !DIExpression()), !dbg [[DBG144]] |
| // CHECK-NEXT: store [10 x [10 x [10 x i32]]] addrspace(1)* [[C]], [10 x [10 x [10 x i32]]] addrspace(1)** [[C_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x [10 x i32]]] addrspace(1)** [[C_ADDR]], metadata [[META146:![0-9]+]], metadata !DIExpression()), !dbg [[DBG147:![0-9]+]] |
| // CHECK-NEXT: store i32 [[A]], i32* [[A_ADDR]], align 4 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32* [[A_ADDR]], metadata [[META148:![0-9]+]], metadata !DIExpression()), !dbg [[DBG149:![0-9]+]] |
| // CHECK-NEXT: store [10 x [10 x i32]] addrspace(1)* [[B]], [10 x [10 x i32]] addrspace(1)** [[B_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x i32]] addrspace(1)** [[B_ADDR]], metadata [[META150:![0-9]+]], metadata !DIExpression()), !dbg [[DBG151:![0-9]+]] |
| // CHECK-NEXT: store i8 addrspace(1)* [[BB]], i8 addrspace(1)** [[BB_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i8 addrspace(1)** [[BB_ADDR]], metadata [[META152:![0-9]+]], metadata !DIExpression()), !dbg [[DBG153:![0-9]+]] |
| // CHECK-NEXT: [[TMP0:%.*]] = load [10 x [10 x [10 x i32]]] addrspace(1)*, [10 x [10 x [10 x i32]]] addrspace(1)** [[C_ADDR]], align 8, !dbg [[DBG154:![0-9]+]] |
| // CHECK-NEXT: [[TMP1:%.*]] = addrspacecast [10 x [10 x [10 x i32]]] addrspace(1)* [[TMP0]] to [10 x [10 x [10 x i32]]]*, !dbg [[DBG154]] |
| // CHECK-NEXT: store [10 x [10 x [10 x i32]]]* [[TMP1]], [10 x [10 x [10 x i32]]]** [[TMP]], align 8, !dbg [[DBG154]] |
| // CHECK-NEXT: [[TMP2:%.*]] = load [10 x [10 x [10 x i32]]]*, [10 x [10 x [10 x i32]]]** [[TMP]], align 8, !dbg [[DBG154]] |
| // CHECK-NEXT: [[TMP3:%.*]] = load [10 x [10 x i32]] addrspace(1)*, [10 x [10 x i32]] addrspace(1)** [[B_ADDR]], align 8, !dbg [[DBG154]] |
| // CHECK-NEXT: [[TMP4:%.*]] = addrspacecast [10 x [10 x i32]] addrspace(1)* [[TMP3]] to [10 x [10 x i32]]*, !dbg [[DBG154]] |
| // CHECK-NEXT: store [10 x [10 x i32]]* [[TMP4]], [10 x [10 x i32]]** [[_TMP1]], align 8, !dbg [[DBG154]] |
| // CHECK-NEXT: [[TMP5:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[_TMP1]], align 8, !dbg [[DBG154]] |
| // CHECK-NEXT: [[TMP6:%.*]] = load i8 addrspace(1)*, i8 addrspace(1)** [[BB_ADDR]], align 8, !dbg [[DBG154]] |
| // CHECK-NEXT: [[TMP7:%.*]] = addrspacecast i8 addrspace(1)* [[TMP6]] to i8*, !dbg [[DBG154]] |
| // CHECK-NEXT: store i8* [[TMP7]], i8** [[_TMP2]], align 8, !dbg [[DBG154]] |
| // CHECK-NEXT: [[TMP8:%.*]] = load i8*, i8** [[_TMP2]], align 8, !dbg [[DBG154]] |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[F]], metadata [[META155:![0-9]+]], metadata !DIExpression()), !dbg [[DBG157:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x [10 x [10 x i32]]], [10 x [10 x [10 x i32]]]* [[TMP2]], i64 0, i64 1, !dbg [[DBG158:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX3:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[ARRAYIDX]], i64 0, i64 1, !dbg [[DBG158]] |
| // CHECK-NEXT: [[ARRAYIDX4:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX3]], i64 0, i64 1, !dbg [[DBG158]] |
| // CHECK-NEXT: store i32* [[ARRAYIDX4]], i32** [[F]], align 8, !dbg [[DBG157]] |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[G]], metadata [[META159:![0-9]+]], metadata !DIExpression()), !dbg [[DBG160:![0-9]+]] |
| // CHECK-NEXT: store i32* [[A_ADDR]], i32** [[G]], align 8, !dbg [[DBG160]] |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[H]], metadata [[META161:![0-9]+]], metadata !DIExpression()), !dbg [[DBG162:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX5:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[TMP5]], i64 0, i64 1, !dbg [[DBG163:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX6:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX5]], i64 0, i64 1, !dbg [[DBG163]] |
| // CHECK-NEXT: store i32* [[ARRAYIDX6]], i32** [[H]], align 8, !dbg [[DBG162]] |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32* [[D]], metadata [[META164:![0-9]+]], metadata !DIExpression()), !dbg [[DBG165:![0-9]+]] |
| // CHECK-NEXT: store i32 15, i32* [[D]], align 4, !dbg [[DBG165]] |
| // CHECK-NEXT: store i32 5, i32* [[A_ADDR]], align 4, !dbg [[DBG166:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX7:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[TMP5]], i64 0, i64 0, !dbg [[DBG167:![0-9]+]] |
| // CHECK-NEXT: [[TMP9:%.*]] = load i32, i32* [[A_ADDR]], align 4, !dbg [[DBG168:![0-9]+]] |
| // CHECK-NEXT: [[IDXPROM:%.*]] = sext i32 [[TMP9]] to i64, !dbg [[DBG167]] |
| // CHECK-NEXT: [[ARRAYIDX8:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX7]], i64 0, i64 [[IDXPROM]], !dbg [[DBG167]] |
| // CHECK-NEXT: store i32 10, i32* [[ARRAYIDX8]], align 4, !dbg [[DBG169:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX9:%.*]] = getelementptr inbounds [10 x [10 x [10 x i32]]], [10 x [10 x [10 x i32]]]* [[TMP2]], i64 0, i64 0, !dbg [[DBG170:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX10:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[ARRAYIDX9]], i64 0, i64 0, !dbg [[DBG170]] |
| // CHECK-NEXT: [[TMP10:%.*]] = load i32, i32* [[A_ADDR]], align 4, !dbg [[DBG171:![0-9]+]] |
| // CHECK-NEXT: [[IDXPROM11:%.*]] = sext i32 [[TMP10]] to i64, !dbg [[DBG170]] |
| // CHECK-NEXT: [[ARRAYIDX12:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX10]], i64 0, i64 [[IDXPROM11]], !dbg [[DBG170]] |
| // CHECK-NEXT: store i32 11, i32* [[ARRAYIDX12]], align 4, !dbg [[DBG172:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX13:%.*]] = getelementptr inbounds [10 x [10 x [10 x i32]]], [10 x [10 x [10 x i32]]]* [[TMP2]], i64 0, i64 0, !dbg [[DBG173:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX14:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[ARRAYIDX13]], i64 0, i64 0, !dbg [[DBG173]] |
| // CHECK-NEXT: [[TMP11:%.*]] = load i32, i32* [[A_ADDR]], align 4, !dbg [[DBG174:![0-9]+]] |
| // CHECK-NEXT: [[IDXPROM15:%.*]] = sext i32 [[TMP11]] to i64, !dbg [[DBG173]] |
| // CHECK-NEXT: [[ARRAYIDX16:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX14]], i64 0, i64 [[IDXPROM15]], !dbg [[DBG173]] |
| // CHECK-NEXT: [[TMP12:%.*]] = load i32, i32* [[ARRAYIDX16]], align 4, !dbg [[DBG173]] |
| // CHECK-NEXT: [[ARRAYIDX17:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[TMP5]], i64 0, i64 0, !dbg [[DBG175:![0-9]+]] |
| // CHECK-NEXT: [[TMP13:%.*]] = load i32, i32* [[A_ADDR]], align 4, !dbg [[DBG176:![0-9]+]] |
| // CHECK-NEXT: [[IDXPROM18:%.*]] = sext i32 [[TMP13]] to i64, !dbg [[DBG175]] |
| // CHECK-NEXT: [[ARRAYIDX19:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX17]], i64 0, i64 [[IDXPROM18]], !dbg [[DBG175]] |
| // CHECK-NEXT: store i32 [[TMP12]], i32* [[ARRAYIDX19]], align 4, !dbg [[DBG177:![0-9]+]] |
| // CHECK-NEXT: [[TMP14:%.*]] = load i8, i8* [[TMP8]], align 1, !dbg [[DBG178:![0-9]+]] |
| // CHECK-NEXT: [[TOBOOL:%.*]] = trunc i8 [[TMP14]] to i1, !dbg [[DBG178]] |
| // CHECK-NEXT: [[CONV:%.*]] = zext i1 [[TOBOOL]] to i32, !dbg [[DBG178]] |
| // CHECK-NEXT: store i32 [[CONV]], i32* [[D]], align 4, !dbg [[DBG179:![0-9]+]] |
| // CHECK-NEXT: ret void, !dbg [[DBG180:![0-9]+]] |
| // |
| // |
| // CHECK-LABEL: define {{[^@]+}}@__omp_outlined__2 |
| // CHECK-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], [10 x [10 x [10 x i32]]]* nonnull align 4 dereferenceable(4000) [[C:%.*]], i64 [[A:%.*]], [10 x [10 x i32]]* nonnull align 4 dereferenceable(400) [[B:%.*]], i8* nonnull align 1 dereferenceable(1) [[BB:%.*]]) #[[ATTR0]] !dbg [[DBG181:![0-9]+]] { |
| // CHECK-NEXT: entry: |
| // CHECK-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[C_ADDR:%.*]] = alloca [10 x [10 x [10 x i32]]]*, align 8 |
| // CHECK-NEXT: [[A_ADDR:%.*]] = alloca i64, align 8 |
| // CHECK-NEXT: [[B_ADDR:%.*]] = alloca [10 x [10 x i32]]*, align 8 |
| // CHECK-NEXT: [[BB_ADDR:%.*]] = alloca i8*, align 8 |
| // CHECK-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[DOTGLOBAL_TID__ADDR]], metadata [[META182:![0-9]+]], metadata !DIExpression()), !dbg [[DBG183:![0-9]+]] |
| // CHECK-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[DOTBOUND_TID__ADDR]], metadata [[META184:![0-9]+]], metadata !DIExpression()), !dbg [[DBG183]] |
| // CHECK-NEXT: store [10 x [10 x [10 x i32]]]* [[C]], [10 x [10 x [10 x i32]]]** [[C_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x [10 x i32]]]** [[C_ADDR]], metadata [[META185:![0-9]+]], metadata !DIExpression()), !dbg [[DBG183]] |
| // CHECK-NEXT: store i64 [[A]], i64* [[A_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i64* [[A_ADDR]], metadata [[META186:![0-9]+]], metadata !DIExpression()), !dbg [[DBG183]] |
| // CHECK-NEXT: store [10 x [10 x i32]]* [[B]], [10 x [10 x i32]]** [[B_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x i32]]** [[B_ADDR]], metadata [[META187:![0-9]+]], metadata !DIExpression()), !dbg [[DBG183]] |
| // CHECK-NEXT: store i8* [[BB]], i8** [[BB_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i8** [[BB_ADDR]], metadata [[META188:![0-9]+]], metadata !DIExpression()), !dbg [[DBG183]] |
| // CHECK-NEXT: [[TMP0:%.*]] = load [10 x [10 x [10 x i32]]]*, [10 x [10 x [10 x i32]]]** [[C_ADDR]], align 8, !dbg [[DBG189:![0-9]+]] |
| // CHECK-NEXT: [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*, !dbg [[DBG189]] |
| // CHECK-NEXT: [[TMP1:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[B_ADDR]], align 8, !dbg [[DBG189]] |
| // CHECK-NEXT: [[TMP2:%.*]] = load i8*, i8** [[BB_ADDR]], align 8, !dbg [[DBG189]] |
| // CHECK-NEXT: [[TMP3:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8, !dbg [[DBG189]] |
| // CHECK-NEXT: [[TMP4:%.*]] = load i32*, i32** [[DOTBOUND_TID__ADDR]], align 8, !dbg [[DBG189]] |
| // CHECK-NEXT: [[TMP5:%.*]] = load [10 x [10 x [10 x i32]]]*, [10 x [10 x [10 x i32]]]** [[C_ADDR]], align 8, !dbg [[DBG189]] |
| // CHECK-NEXT: [[TMP6:%.*]] = load i32, i32* [[CONV]], align 8, !dbg [[DBG189]] |
| // CHECK-NEXT: [[TMP7:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[B_ADDR]], align 8, !dbg [[DBG189]] |
| // CHECK-NEXT: [[TMP8:%.*]] = load i8*, i8** [[BB_ADDR]], align 8, !dbg [[DBG189]] |
| // CHECK-NEXT: [[TMP9:%.*]] = addrspacecast [10 x [10 x [10 x i32]]]* [[TMP5]] to [10 x [10 x [10 x i32]]] addrspace(1)*, !dbg [[DBG189]] |
| // CHECK-NEXT: [[TMP10:%.*]] = addrspacecast [10 x [10 x i32]]* [[TMP7]] to [10 x [10 x i32]] addrspace(1)*, !dbg [[DBG189]] |
| // CHECK-NEXT: [[TMP11:%.*]] = addrspacecast i8* [[TMP8]] to i8 addrspace(1)*, !dbg [[DBG189]] |
| // CHECK-NEXT: call void @__omp_outlined___debug__1(i32* [[TMP3]], i32* [[TMP4]], [10 x [10 x [10 x i32]]] addrspace(1)* [[TMP9]], i32 [[TMP6]], [10 x [10 x i32]] addrspace(1)* [[TMP10]], i8 addrspace(1)* [[TMP11]]) #[[ATTR4]], !dbg [[DBG189]] |
| // CHECK-NEXT: ret void, !dbg [[DBG189]] |
| // |
| // |
| // CHECK-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l37 |
| // CHECK-SAME: ([10 x [10 x [10 x i32]]]* nonnull align 4 dereferenceable(4000) [[C:%.*]], i64 [[A:%.*]], [10 x [10 x i32]]* nonnull align 4 dereferenceable(400) [[B:%.*]], i8* nonnull align 1 dereferenceable(1) [[BB:%.*]]) #[[ATTR0]] !dbg [[DBG190:![0-9]+]] { |
| // CHECK-NEXT: entry: |
| // CHECK-NEXT: [[C_ADDR:%.*]] = alloca [10 x [10 x [10 x i32]]]*, align 8 |
| // CHECK-NEXT: [[A_ADDR:%.*]] = alloca i64, align 8 |
| // CHECK-NEXT: [[B_ADDR:%.*]] = alloca [10 x [10 x i32]]*, align 8 |
| // CHECK-NEXT: [[BB_ADDR:%.*]] = alloca i8*, align 8 |
| // CHECK-NEXT: store [10 x [10 x [10 x i32]]]* [[C]], [10 x [10 x [10 x i32]]]** [[C_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x [10 x i32]]]** [[C_ADDR]], metadata [[META191:![0-9]+]], metadata !DIExpression()), !dbg [[DBG192:![0-9]+]] |
| // CHECK-NEXT: store i64 [[A]], i64* [[A_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i64* [[A_ADDR]], metadata [[META193:![0-9]+]], metadata !DIExpression()), !dbg [[DBG192]] |
| // CHECK-NEXT: store [10 x [10 x i32]]* [[B]], [10 x [10 x i32]]** [[B_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x i32]]** [[B_ADDR]], metadata [[META194:![0-9]+]], metadata !DIExpression()), !dbg [[DBG192]] |
| // CHECK-NEXT: store i8* [[BB]], i8** [[BB_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i8** [[BB_ADDR]], metadata [[META195:![0-9]+]], metadata !DIExpression()), !dbg [[DBG192]] |
| // CHECK-NEXT: [[TMP0:%.*]] = load [10 x [10 x [10 x i32]]]*, [10 x [10 x [10 x i32]]]** [[C_ADDR]], align 8, !dbg [[DBG196:![0-9]+]] |
| // CHECK-NEXT: [[CONV:%.*]] = bitcast i64* [[A_ADDR]] to i32*, !dbg [[DBG196]] |
| // CHECK-NEXT: [[TMP1:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[B_ADDR]], align 8, !dbg [[DBG196]] |
| // CHECK-NEXT: [[TMP2:%.*]] = load i8*, i8** [[BB_ADDR]], align 8, !dbg [[DBG196]] |
| // CHECK-NEXT: [[TMP3:%.*]] = load [10 x [10 x [10 x i32]]]*, [10 x [10 x [10 x i32]]]** [[C_ADDR]], align 8, !dbg [[DBG196]] |
| // CHECK-NEXT: [[TMP4:%.*]] = load i32, i32* [[CONV]], align 8, !dbg [[DBG196]] |
| // CHECK-NEXT: [[TMP5:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[B_ADDR]], align 8, !dbg [[DBG196]] |
| // CHECK-NEXT: [[TMP6:%.*]] = load i8*, i8** [[BB_ADDR]], align 8, !dbg [[DBG196]] |
| // CHECK-NEXT: [[TMP7:%.*]] = addrspacecast [10 x [10 x [10 x i32]]]* [[TMP3]] to [10 x [10 x [10 x i32]]] addrspace(1)*, !dbg [[DBG196]] |
| // CHECK-NEXT: [[TMP8:%.*]] = addrspacecast [10 x [10 x i32]]* [[TMP5]] to [10 x [10 x i32]] addrspace(1)*, !dbg [[DBG196]] |
| // CHECK-NEXT: [[TMP9:%.*]] = addrspacecast i8* [[TMP6]] to i8 addrspace(1)*, !dbg [[DBG196]] |
| // CHECK-NEXT: call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l37_debug__([10 x [10 x [10 x i32]]] addrspace(1)* [[TMP7]], i32 [[TMP4]], [10 x [10 x i32]] addrspace(1)* [[TMP8]], i8 addrspace(1)* [[TMP9]]) #[[ATTR4]], !dbg [[DBG196]] |
| // CHECK-NEXT: ret void, !dbg [[DBG196]] |
| // |
| // |
| // CHECK-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l51_debug__ |
| // CHECK-SAME: ([10 x [10 x [10 x i32]]] addrspace(1)* noalias [[C:%.*]], i32 addrspace(1)* noalias [[A:%.*]], [10 x [10 x i32]] addrspace(1)* noalias [[B:%.*]], i8 addrspace(1)* noalias [[BB:%.*]]) #[[ATTR0]] !dbg [[DBG197:![0-9]+]] { |
| // CHECK-NEXT: entry: |
| // CHECK-NEXT: [[C_ADDR:%.*]] = alloca [10 x [10 x [10 x i32]]] addrspace(1)*, align 8 |
| // CHECK-NEXT: [[A_ADDR:%.*]] = alloca i32 addrspace(1)*, align 8 |
| // CHECK-NEXT: [[B_ADDR:%.*]] = alloca [10 x [10 x i32]] addrspace(1)*, align 8 |
| // CHECK-NEXT: [[BB_ADDR:%.*]] = alloca i8 addrspace(1)*, align 8 |
| // CHECK-NEXT: [[TMP:%.*]] = alloca [10 x [10 x [10 x i32]]]*, align 8 |
| // CHECK-NEXT: [[_TMP1:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[_TMP2:%.*]] = alloca [10 x [10 x i32]]*, align 8 |
| // CHECK-NEXT: [[_TMP3:%.*]] = alloca i8*, align 8 |
| // CHECK-NEXT: [[CAPTURED_VARS_ADDRS:%.*]] = alloca [4 x i8*], align 8 |
| // CHECK-NEXT: store [10 x [10 x [10 x i32]]] addrspace(1)* [[C]], [10 x [10 x [10 x i32]]] addrspace(1)** [[C_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x [10 x i32]]] addrspace(1)** [[C_ADDR]], metadata [[META202:![0-9]+]], metadata !DIExpression()), !dbg [[DBG203:![0-9]+]] |
| // CHECK-NEXT: store i32 addrspace(1)* [[A]], i32 addrspace(1)** [[A_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32 addrspace(1)** [[A_ADDR]], metadata [[META204:![0-9]+]], metadata !DIExpression()), !dbg [[DBG205:![0-9]+]] |
| // CHECK-NEXT: store [10 x [10 x i32]] addrspace(1)* [[B]], [10 x [10 x i32]] addrspace(1)** [[B_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x i32]] addrspace(1)** [[B_ADDR]], metadata [[META206:![0-9]+]], metadata !DIExpression()), !dbg [[DBG207:![0-9]+]] |
| // CHECK-NEXT: store i8 addrspace(1)* [[BB]], i8 addrspace(1)** [[BB_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i8 addrspace(1)** [[BB_ADDR]], metadata [[META208:![0-9]+]], metadata !DIExpression()), !dbg [[DBG209:![0-9]+]] |
| // CHECK-NEXT: [[TMP0:%.*]] = load [10 x [10 x [10 x i32]]] addrspace(1)*, [10 x [10 x [10 x i32]]] addrspace(1)** [[C_ADDR]], align 8, !dbg [[DBG210:![0-9]+]] |
| // CHECK-NEXT: [[TMP1:%.*]] = addrspacecast [10 x [10 x [10 x i32]]] addrspace(1)* [[TMP0]] to [10 x [10 x [10 x i32]]]*, !dbg [[DBG210]] |
| // CHECK-NEXT: store [10 x [10 x [10 x i32]]]* [[TMP1]], [10 x [10 x [10 x i32]]]** [[TMP]], align 8, !dbg [[DBG210]] |
| // CHECK-NEXT: [[TMP2:%.*]] = load [10 x [10 x [10 x i32]]]*, [10 x [10 x [10 x i32]]]** [[TMP]], align 8, !dbg [[DBG210]] |
| // CHECK-NEXT: [[TMP3:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)** [[A_ADDR]], align 8, !dbg [[DBG210]] |
| // CHECK-NEXT: [[TMP4:%.*]] = addrspacecast i32 addrspace(1)* [[TMP3]] to i32*, !dbg [[DBG210]] |
| // CHECK-NEXT: store i32* [[TMP4]], i32** [[_TMP1]], align 8, !dbg [[DBG210]] |
| // CHECK-NEXT: [[TMP5:%.*]] = load i32*, i32** [[_TMP1]], align 8, !dbg [[DBG210]] |
| // CHECK-NEXT: [[TMP6:%.*]] = load [10 x [10 x i32]] addrspace(1)*, [10 x [10 x i32]] addrspace(1)** [[B_ADDR]], align 8, !dbg [[DBG210]] |
| // CHECK-NEXT: [[TMP7:%.*]] = addrspacecast [10 x [10 x i32]] addrspace(1)* [[TMP6]] to [10 x [10 x i32]]*, !dbg [[DBG210]] |
| // CHECK-NEXT: store [10 x [10 x i32]]* [[TMP7]], [10 x [10 x i32]]** [[_TMP2]], align 8, !dbg [[DBG210]] |
| // CHECK-NEXT: [[TMP8:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[_TMP2]], align 8, !dbg [[DBG210]] |
| // CHECK-NEXT: [[TMP9:%.*]] = load i8 addrspace(1)*, i8 addrspace(1)** [[BB_ADDR]], align 8, !dbg [[DBG210]] |
| // CHECK-NEXT: [[TMP10:%.*]] = addrspacecast i8 addrspace(1)* [[TMP9]] to i8*, !dbg [[DBG210]] |
| // CHECK-NEXT: store i8* [[TMP10]], i8** [[_TMP3]], align 8, !dbg [[DBG210]] |
| // CHECK-NEXT: [[TMP11:%.*]] = load i8*, i8** [[_TMP3]], align 8, !dbg [[DBG210]] |
| // CHECK-NEXT: [[NVPTX_NUM_THREADS:%.*]] = call i32 @llvm.nvvm.read.ptx.sreg.ntid.x(), !dbg [[DBG210]] |
| // CHECK-NEXT: call void @__kmpc_spmd_kernel_init(i32 [[NVPTX_NUM_THREADS]], i16 1), !dbg [[DBG210]] |
| // CHECK-NEXT: call void @__kmpc_data_sharing_init_stack_spmd(), !dbg [[DBG210]] |
| // CHECK-NEXT: br label [[DOTEXECUTE:%.*]], !dbg [[DBG210]] |
| // CHECK: .execute: |
| // CHECK-NEXT: [[TMP12:%.*]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* @[[GLOB5:[0-9]+]]) |
| // CHECK-NEXT: [[TMP13:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 0, !dbg [[DBG211:![0-9]+]] |
| // CHECK-NEXT: [[TMP14:%.*]] = bitcast [10 x [10 x [10 x i32]]]* [[TMP2]] to i8*, !dbg [[DBG211]] |
| // CHECK-NEXT: store i8* [[TMP14]], i8** [[TMP13]], align 8, !dbg [[DBG211]] |
| // CHECK-NEXT: [[TMP15:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 1, !dbg [[DBG211]] |
| // CHECK-NEXT: [[TMP16:%.*]] = bitcast i32* [[TMP5]] to i8*, !dbg [[DBG211]] |
| // CHECK-NEXT: store i8* [[TMP16]], i8** [[TMP15]], align 8, !dbg [[DBG211]] |
| // CHECK-NEXT: [[TMP17:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 2, !dbg [[DBG211]] |
| // CHECK-NEXT: [[TMP18:%.*]] = bitcast [10 x [10 x i32]]* [[TMP8]] to i8*, !dbg [[DBG211]] |
| // CHECK-NEXT: store i8* [[TMP18]], i8** [[TMP17]], align 8, !dbg [[DBG211]] |
| // CHECK-NEXT: [[TMP19:%.*]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[CAPTURED_VARS_ADDRS]], i64 0, i64 3, !dbg [[DBG211]] |
| // CHECK-NEXT: store i8* [[TMP11]], i8** [[TMP19]], align 8, !dbg [[DBG211]] |
| // CHECK-NEXT: [[TMP20:%.*]] = bitcast [4 x i8*]* [[CAPTURED_VARS_ADDRS]] to i8**, !dbg [[DBG211]] |
| // CHECK-NEXT: call void @__kmpc_parallel_51(%struct.ident_t* @[[GLOB5]], i32 [[TMP12]], i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*, [10 x [10 x [10 x i32]]]*, i32*, [10 x [10 x i32]]*, i8*)* @__omp_outlined__4 to i8*), i8* null, i8** [[TMP20]], i64 4), !dbg [[DBG211]] |
| // CHECK-NEXT: br label [[DOTOMP_DEINIT:%.*]], !dbg [[DBG212:![0-9]+]] |
| // CHECK: .omp.deinit: |
| // CHECK-NEXT: call void @__kmpc_spmd_kernel_deinit_v2(i16 1), !dbg [[DBG212]] |
| // CHECK-NEXT: br label [[DOTEXIT:%.*]], !dbg [[DBG212]] |
| // CHECK: .exit: |
| // CHECK-NEXT: ret void, !dbg [[DBG214:![0-9]+]] |
| // |
| // |
| // CHECK-LABEL: define {{[^@]+}}@__omp_outlined___debug__3 |
| // CHECK-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], [10 x [10 x [10 x i32]]] addrspace(1)* noalias [[C:%.*]], i32 addrspace(1)* noalias [[A:%.*]], [10 x [10 x i32]] addrspace(1)* noalias [[B:%.*]], i8 addrspace(1)* noalias [[BB:%.*]]) #[[ATTR0]] !dbg [[DBG215:![0-9]+]] { |
| // CHECK-NEXT: entry: |
| // CHECK-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[C_ADDR:%.*]] = alloca [10 x [10 x [10 x i32]]] addrspace(1)*, align 8 |
| // CHECK-NEXT: [[A_ADDR:%.*]] = alloca i32 addrspace(1)*, align 8 |
| // CHECK-NEXT: [[B_ADDR:%.*]] = alloca [10 x [10 x i32]] addrspace(1)*, align 8 |
| // CHECK-NEXT: [[BB_ADDR:%.*]] = alloca i8 addrspace(1)*, align 8 |
| // CHECK-NEXT: [[TMP:%.*]] = alloca [10 x [10 x [10 x i32]]]*, align 8 |
| // CHECK-NEXT: [[_TMP1:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[_TMP2:%.*]] = alloca [10 x [10 x i32]]*, align 8 |
| // CHECK-NEXT: [[_TMP3:%.*]] = alloca i8*, align 8 |
| // CHECK-NEXT: [[F:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[G:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[H:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[D:%.*]] = alloca i32, align 4 |
| // CHECK-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[DOTGLOBAL_TID__ADDR]], metadata [[META218:![0-9]+]], metadata !DIExpression()), !dbg [[DBG219:![0-9]+]] |
| // CHECK-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[DOTBOUND_TID__ADDR]], metadata [[META220:![0-9]+]], metadata !DIExpression()), !dbg [[DBG219]] |
| // CHECK-NEXT: store [10 x [10 x [10 x i32]]] addrspace(1)* [[C]], [10 x [10 x [10 x i32]]] addrspace(1)** [[C_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x [10 x i32]]] addrspace(1)** [[C_ADDR]], metadata [[META221:![0-9]+]], metadata !DIExpression()), !dbg [[DBG222:![0-9]+]] |
| // CHECK-NEXT: store i32 addrspace(1)* [[A]], i32 addrspace(1)** [[A_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32 addrspace(1)** [[A_ADDR]], metadata [[META223:![0-9]+]], metadata !DIExpression()), !dbg [[DBG224:![0-9]+]] |
| // CHECK-NEXT: store [10 x [10 x i32]] addrspace(1)* [[B]], [10 x [10 x i32]] addrspace(1)** [[B_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x i32]] addrspace(1)** [[B_ADDR]], metadata [[META225:![0-9]+]], metadata !DIExpression()), !dbg [[DBG226:![0-9]+]] |
| // CHECK-NEXT: store i8 addrspace(1)* [[BB]], i8 addrspace(1)** [[BB_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i8 addrspace(1)** [[BB_ADDR]], metadata [[META227:![0-9]+]], metadata !DIExpression()), !dbg [[DBG228:![0-9]+]] |
| // CHECK-NEXT: [[TMP0:%.*]] = load [10 x [10 x [10 x i32]]] addrspace(1)*, [10 x [10 x [10 x i32]]] addrspace(1)** [[C_ADDR]], align 8, !dbg [[DBG229:![0-9]+]] |
| // CHECK-NEXT: [[TMP1:%.*]] = addrspacecast [10 x [10 x [10 x i32]]] addrspace(1)* [[TMP0]] to [10 x [10 x [10 x i32]]]*, !dbg [[DBG229]] |
| // CHECK-NEXT: store [10 x [10 x [10 x i32]]]* [[TMP1]], [10 x [10 x [10 x i32]]]** [[TMP]], align 8, !dbg [[DBG229]] |
| // CHECK-NEXT: [[TMP2:%.*]] = load [10 x [10 x [10 x i32]]]*, [10 x [10 x [10 x i32]]]** [[TMP]], align 8, !dbg [[DBG229]] |
| // CHECK-NEXT: [[TMP3:%.*]] = load i32 addrspace(1)*, i32 addrspace(1)** [[A_ADDR]], align 8, !dbg [[DBG229]] |
| // CHECK-NEXT: [[TMP4:%.*]] = addrspacecast i32 addrspace(1)* [[TMP3]] to i32*, !dbg [[DBG229]] |
| // CHECK-NEXT: store i32* [[TMP4]], i32** [[_TMP1]], align 8, !dbg [[DBG229]] |
| // CHECK-NEXT: [[TMP5:%.*]] = load i32*, i32** [[_TMP1]], align 8, !dbg [[DBG229]] |
| // CHECK-NEXT: [[TMP6:%.*]] = load [10 x [10 x i32]] addrspace(1)*, [10 x [10 x i32]] addrspace(1)** [[B_ADDR]], align 8, !dbg [[DBG229]] |
| // CHECK-NEXT: [[TMP7:%.*]] = addrspacecast [10 x [10 x i32]] addrspace(1)* [[TMP6]] to [10 x [10 x i32]]*, !dbg [[DBG229]] |
| // CHECK-NEXT: store [10 x [10 x i32]]* [[TMP7]], [10 x [10 x i32]]** [[_TMP2]], align 8, !dbg [[DBG229]] |
| // CHECK-NEXT: [[TMP8:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[_TMP2]], align 8, !dbg [[DBG229]] |
| // CHECK-NEXT: [[TMP9:%.*]] = load i8 addrspace(1)*, i8 addrspace(1)** [[BB_ADDR]], align 8, !dbg [[DBG229]] |
| // CHECK-NEXT: [[TMP10:%.*]] = addrspacecast i8 addrspace(1)* [[TMP9]] to i8*, !dbg [[DBG229]] |
| // CHECK-NEXT: store i8* [[TMP10]], i8** [[_TMP3]], align 8, !dbg [[DBG229]] |
| // CHECK-NEXT: [[TMP11:%.*]] = load i8*, i8** [[_TMP3]], align 8, !dbg [[DBG229]] |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[F]], metadata [[META230:![0-9]+]], metadata !DIExpression()), !dbg [[DBG232:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [10 x [10 x [10 x i32]]], [10 x [10 x [10 x i32]]]* [[TMP2]], i64 0, i64 1, !dbg [[DBG233:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX4:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[ARRAYIDX]], i64 0, i64 1, !dbg [[DBG233]] |
| // CHECK-NEXT: [[ARRAYIDX5:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX4]], i64 0, i64 1, !dbg [[DBG233]] |
| // CHECK-NEXT: store i32* [[ARRAYIDX5]], i32** [[F]], align 8, !dbg [[DBG232]] |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[G]], metadata [[META234:![0-9]+]], metadata !DIExpression()), !dbg [[DBG235:![0-9]+]] |
| // CHECK-NEXT: store i32* [[TMP5]], i32** [[G]], align 8, !dbg [[DBG235]] |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[H]], metadata [[META236:![0-9]+]], metadata !DIExpression()), !dbg [[DBG237:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX6:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[TMP8]], i64 0, i64 1, !dbg [[DBG238:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX7:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX6]], i64 0, i64 1, !dbg [[DBG238]] |
| // CHECK-NEXT: store i32* [[ARRAYIDX7]], i32** [[H]], align 8, !dbg [[DBG237]] |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32* [[D]], metadata [[META239:![0-9]+]], metadata !DIExpression()), !dbg [[DBG240:![0-9]+]] |
| // CHECK-NEXT: store i32 15, i32* [[D]], align 4, !dbg [[DBG240]] |
| // CHECK-NEXT: store i32 5, i32* [[TMP5]], align 4, !dbg [[DBG241:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX8:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[TMP8]], i64 0, i64 0, !dbg [[DBG242:![0-9]+]] |
| // CHECK-NEXT: [[TMP12:%.*]] = load i32, i32* [[TMP5]], align 4, !dbg [[DBG243:![0-9]+]] |
| // CHECK-NEXT: [[IDXPROM:%.*]] = sext i32 [[TMP12]] to i64, !dbg [[DBG242]] |
| // CHECK-NEXT: [[ARRAYIDX9:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX8]], i64 0, i64 [[IDXPROM]], !dbg [[DBG242]] |
| // CHECK-NEXT: store i32 10, i32* [[ARRAYIDX9]], align 4, !dbg [[DBG244:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX10:%.*]] = getelementptr inbounds [10 x [10 x [10 x i32]]], [10 x [10 x [10 x i32]]]* [[TMP2]], i64 0, i64 0, !dbg [[DBG245:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX11:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[ARRAYIDX10]], i64 0, i64 0, !dbg [[DBG245]] |
| // CHECK-NEXT: [[TMP13:%.*]] = load i32, i32* [[TMP5]], align 4, !dbg [[DBG246:![0-9]+]] |
| // CHECK-NEXT: [[IDXPROM12:%.*]] = sext i32 [[TMP13]] to i64, !dbg [[DBG245]] |
| // CHECK-NEXT: [[ARRAYIDX13:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX11]], i64 0, i64 [[IDXPROM12]], !dbg [[DBG245]] |
| // CHECK-NEXT: store i32 11, i32* [[ARRAYIDX13]], align 4, !dbg [[DBG247:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX14:%.*]] = getelementptr inbounds [10 x [10 x [10 x i32]]], [10 x [10 x [10 x i32]]]* [[TMP2]], i64 0, i64 0, !dbg [[DBG248:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX15:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[ARRAYIDX14]], i64 0, i64 0, !dbg [[DBG248]] |
| // CHECK-NEXT: [[TMP14:%.*]] = load i32, i32* [[TMP5]], align 4, !dbg [[DBG249:![0-9]+]] |
| // CHECK-NEXT: [[IDXPROM16:%.*]] = sext i32 [[TMP14]] to i64, !dbg [[DBG248]] |
| // CHECK-NEXT: [[ARRAYIDX17:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX15]], i64 0, i64 [[IDXPROM16]], !dbg [[DBG248]] |
| // CHECK-NEXT: [[TMP15:%.*]] = load i32, i32* [[ARRAYIDX17]], align 4, !dbg [[DBG248]] |
| // CHECK-NEXT: [[ARRAYIDX18:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[TMP8]], i64 0, i64 0, !dbg [[DBG250:![0-9]+]] |
| // CHECK-NEXT: [[TMP16:%.*]] = load i32, i32* [[TMP5]], align 4, !dbg [[DBG251:![0-9]+]] |
| // CHECK-NEXT: [[IDXPROM19:%.*]] = sext i32 [[TMP16]] to i64, !dbg [[DBG250]] |
| // CHECK-NEXT: [[ARRAYIDX20:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX18]], i64 0, i64 [[IDXPROM19]], !dbg [[DBG250]] |
| // CHECK-NEXT: store i32 [[TMP15]], i32* [[ARRAYIDX20]], align 4, !dbg [[DBG252:![0-9]+]] |
| // CHECK-NEXT: [[ARRAYIDX21:%.*]] = getelementptr inbounds [10 x [10 x i32]], [10 x [10 x i32]]* [[TMP8]], i64 0, i64 0, !dbg [[DBG253:![0-9]+]] |
| // CHECK-NEXT: [[TMP17:%.*]] = load i32, i32* [[TMP5]], align 4, !dbg [[DBG254:![0-9]+]] |
| // CHECK-NEXT: [[IDXPROM22:%.*]] = sext i32 [[TMP17]] to i64, !dbg [[DBG253]] |
| // CHECK-NEXT: [[ARRAYIDX23:%.*]] = getelementptr inbounds [10 x i32], [10 x i32]* [[ARRAYIDX21]], i64 0, i64 [[IDXPROM22]], !dbg [[DBG253]] |
| // CHECK-NEXT: [[TMP18:%.*]] = load i32, i32* [[ARRAYIDX23]], align 4, !dbg [[DBG253]] |
| // CHECK-NEXT: [[TOBOOL:%.*]] = icmp ne i32 [[TMP18]], 0, !dbg [[DBG253]] |
| // CHECK-NEXT: [[FROMBOOL:%.*]] = zext i1 [[TOBOOL]] to i8, !dbg [[DBG255:![0-9]+]] |
| // CHECK-NEXT: store i8 [[FROMBOOL]], i8* [[TMP11]], align 1, !dbg [[DBG255]] |
| // CHECK-NEXT: ret void, !dbg [[DBG256:![0-9]+]] |
| // |
| // |
| // CHECK-LABEL: define {{[^@]+}}@__omp_outlined__4 |
| // CHECK-SAME: (i32* noalias [[DOTGLOBAL_TID_:%.*]], i32* noalias [[DOTBOUND_TID_:%.*]], [10 x [10 x [10 x i32]]]* nonnull align 4 dereferenceable(4000) [[C:%.*]], i32* nonnull align 4 dereferenceable(4) [[A:%.*]], [10 x [10 x i32]]* nonnull align 4 dereferenceable(400) [[B:%.*]], i8* nonnull align 1 dereferenceable(1) [[BB:%.*]]) #[[ATTR0]] !dbg [[DBG257:![0-9]+]] { |
| // CHECK-NEXT: entry: |
| // CHECK-NEXT: [[DOTGLOBAL_TID__ADDR:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[DOTBOUND_TID__ADDR:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[C_ADDR:%.*]] = alloca [10 x [10 x [10 x i32]]]*, align 8 |
| // CHECK-NEXT: [[A_ADDR:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[B_ADDR:%.*]] = alloca [10 x [10 x i32]]*, align 8 |
| // CHECK-NEXT: [[BB_ADDR:%.*]] = alloca i8*, align 8 |
| // CHECK-NEXT: store i32* [[DOTGLOBAL_TID_]], i32** [[DOTGLOBAL_TID__ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[DOTGLOBAL_TID__ADDR]], metadata [[META260:![0-9]+]], metadata !DIExpression()), !dbg [[DBG261:![0-9]+]] |
| // CHECK-NEXT: store i32* [[DOTBOUND_TID_]], i32** [[DOTBOUND_TID__ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[DOTBOUND_TID__ADDR]], metadata [[META262:![0-9]+]], metadata !DIExpression()), !dbg [[DBG261]] |
| // CHECK-NEXT: store [10 x [10 x [10 x i32]]]* [[C]], [10 x [10 x [10 x i32]]]** [[C_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x [10 x i32]]]** [[C_ADDR]], metadata [[META263:![0-9]+]], metadata !DIExpression()), !dbg [[DBG261]] |
| // CHECK-NEXT: store i32* [[A]], i32** [[A_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[A_ADDR]], metadata [[META264:![0-9]+]], metadata !DIExpression()), !dbg [[DBG261]] |
| // CHECK-NEXT: store [10 x [10 x i32]]* [[B]], [10 x [10 x i32]]** [[B_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x i32]]** [[B_ADDR]], metadata [[META265:![0-9]+]], metadata !DIExpression()), !dbg [[DBG261]] |
| // CHECK-NEXT: store i8* [[BB]], i8** [[BB_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i8** [[BB_ADDR]], metadata [[META266:![0-9]+]], metadata !DIExpression()), !dbg [[DBG261]] |
| // CHECK-NEXT: [[TMP0:%.*]] = load [10 x [10 x [10 x i32]]]*, [10 x [10 x [10 x i32]]]** [[C_ADDR]], align 8, !dbg [[DBG267:![0-9]+]] |
| // CHECK-NEXT: [[TMP1:%.*]] = load i32*, i32** [[A_ADDR]], align 8, !dbg [[DBG267]] |
| // CHECK-NEXT: [[TMP2:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[B_ADDR]], align 8, !dbg [[DBG267]] |
| // CHECK-NEXT: [[TMP3:%.*]] = load i8*, i8** [[BB_ADDR]], align 8, !dbg [[DBG267]] |
| // CHECK-NEXT: [[TMP4:%.*]] = load i32*, i32** [[DOTGLOBAL_TID__ADDR]], align 8, !dbg [[DBG267]] |
| // CHECK-NEXT: [[TMP5:%.*]] = load i32*, i32** [[DOTBOUND_TID__ADDR]], align 8, !dbg [[DBG267]] |
| // CHECK-NEXT: [[TMP6:%.*]] = load [10 x [10 x [10 x i32]]]*, [10 x [10 x [10 x i32]]]** [[C_ADDR]], align 8, !dbg [[DBG267]] |
| // CHECK-NEXT: [[TMP7:%.*]] = load i32*, i32** [[A_ADDR]], align 8, !dbg [[DBG267]] |
| // CHECK-NEXT: [[TMP8:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[B_ADDR]], align 8, !dbg [[DBG267]] |
| // CHECK-NEXT: [[TMP9:%.*]] = load i8*, i8** [[BB_ADDR]], align 8, !dbg [[DBG267]] |
| // CHECK-NEXT: [[TMP10:%.*]] = addrspacecast [10 x [10 x [10 x i32]]]* [[TMP6]] to [10 x [10 x [10 x i32]]] addrspace(1)*, !dbg [[DBG267]] |
| // CHECK-NEXT: [[TMP11:%.*]] = addrspacecast i32* [[TMP7]] to i32 addrspace(1)*, !dbg [[DBG267]] |
| // CHECK-NEXT: [[TMP12:%.*]] = addrspacecast [10 x [10 x i32]]* [[TMP8]] to [10 x [10 x i32]] addrspace(1)*, !dbg [[DBG267]] |
| // CHECK-NEXT: [[TMP13:%.*]] = addrspacecast i8* [[TMP9]] to i8 addrspace(1)*, !dbg [[DBG267]] |
| // CHECK-NEXT: call void @__omp_outlined___debug__3(i32* [[TMP4]], i32* [[TMP5]], [10 x [10 x [10 x i32]]] addrspace(1)* [[TMP10]], i32 addrspace(1)* [[TMP11]], [10 x [10 x i32]] addrspace(1)* [[TMP12]], i8 addrspace(1)* [[TMP13]]) #[[ATTR4]], !dbg [[DBG267]] |
| // CHECK-NEXT: ret void, !dbg [[DBG267]] |
| // |
| // |
| // CHECK-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l51 |
| // CHECK-SAME: ([10 x [10 x [10 x i32]]]* nonnull align 4 dereferenceable(4000) [[C:%.*]], i32* nonnull align 4 dereferenceable(4) [[A:%.*]], [10 x [10 x i32]]* nonnull align 4 dereferenceable(400) [[B:%.*]], i8* nonnull align 1 dereferenceable(1) [[BB:%.*]]) #[[ATTR0]] !dbg [[DBG268:![0-9]+]] { |
| // CHECK-NEXT: entry: |
| // CHECK-NEXT: [[C_ADDR:%.*]] = alloca [10 x [10 x [10 x i32]]]*, align 8 |
| // CHECK-NEXT: [[A_ADDR:%.*]] = alloca i32*, align 8 |
| // CHECK-NEXT: [[B_ADDR:%.*]] = alloca [10 x [10 x i32]]*, align 8 |
| // CHECK-NEXT: [[BB_ADDR:%.*]] = alloca i8*, align 8 |
| // CHECK-NEXT: store [10 x [10 x [10 x i32]]]* [[C]], [10 x [10 x [10 x i32]]]** [[C_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x [10 x i32]]]** [[C_ADDR]], metadata [[META271:![0-9]+]], metadata !DIExpression()), !dbg [[DBG272:![0-9]+]] |
| // CHECK-NEXT: store i32* [[A]], i32** [[A_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i32** [[A_ADDR]], metadata [[META273:![0-9]+]], metadata !DIExpression()), !dbg [[DBG272]] |
| // CHECK-NEXT: store [10 x [10 x i32]]* [[B]], [10 x [10 x i32]]** [[B_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata [10 x [10 x i32]]** [[B_ADDR]], metadata [[META274:![0-9]+]], metadata !DIExpression()), !dbg [[DBG272]] |
| // CHECK-NEXT: store i8* [[BB]], i8** [[BB_ADDR]], align 8 |
| // CHECK-NEXT: call void @llvm.dbg.declare(metadata i8** [[BB_ADDR]], metadata [[META275:![0-9]+]], metadata !DIExpression()), !dbg [[DBG272]] |
| // CHECK-NEXT: [[TMP0:%.*]] = load [10 x [10 x [10 x i32]]]*, [10 x [10 x [10 x i32]]]** [[C_ADDR]], align 8, !dbg [[DBG276:![0-9]+]] |
| // CHECK-NEXT: [[TMP1:%.*]] = load i32*, i32** [[A_ADDR]], align 8, !dbg [[DBG276]] |
| // CHECK-NEXT: [[TMP2:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[B_ADDR]], align 8, !dbg [[DBG276]] |
| // CHECK-NEXT: [[TMP3:%.*]] = load i8*, i8** [[BB_ADDR]], align 8, !dbg [[DBG276]] |
| // CHECK-NEXT: [[TMP4:%.*]] = load [10 x [10 x [10 x i32]]]*, [10 x [10 x [10 x i32]]]** [[C_ADDR]], align 8, !dbg [[DBG276]] |
| // CHECK-NEXT: [[TMP5:%.*]] = load i32*, i32** [[A_ADDR]], align 8, !dbg [[DBG276]] |
| // CHECK-NEXT: [[TMP6:%.*]] = load [10 x [10 x i32]]*, [10 x [10 x i32]]** [[B_ADDR]], align 8, !dbg [[DBG276]] |
| // CHECK-NEXT: [[TMP7:%.*]] = load i8*, i8** [[BB_ADDR]], align 8, !dbg [[DBG276]] |
| // CHECK-NEXT: [[TMP8:%.*]] = addrspacecast [10 x [10 x [10 x i32]]]* [[TMP4]] to [10 x [10 x [10 x i32]]] addrspace(1)*, !dbg [[DBG276]] |
| // CHECK-NEXT: [[TMP9:%.*]] = addrspacecast i32* [[TMP5]] to i32 addrspace(1)*, !dbg [[DBG276]] |
| // CHECK-NEXT: [[TMP10:%.*]] = addrspacecast [10 x [10 x i32]]* [[TMP6]] to [10 x [10 x i32]] addrspace(1)*, !dbg [[DBG276]] |
| // CHECK-NEXT: [[TMP11:%.*]] = addrspacecast i8* [[TMP7]] to i8 addrspace(1)*, !dbg [[DBG276]] |
| // CHECK-NEXT: call void @{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_main_l51_debug__([10 x [10 x [10 x i32]]] addrspace(1)* [[TMP8]], i32 addrspace(1)* [[TMP9]], [10 x [10 x i32]] addrspace(1)* [[TMP10]], i8 addrspace(1)* [[TMP11]]) #[[ATTR4]], !dbg [[DBG276]] |
| // CHECK-NEXT: ret void, !dbg [[DBG276]] |
| // |
